The field's beat trackers find the bar the detector cannot, and the one with a bar model earns the automatic re-draft

Listening

Study Overview

The Question

Thirty-five studies in, the program's standing read was that the automatic detector's tempo and meter are the weak link and that a visitor who taps the tempo and picks the meter is the realistic product. Kevin rejected that framing and asked for a research pass on what had not been tried. The resulting proposal reframed the target: the product needs the bar (its period, its grouping, and where it starts), not a tempo number, and the field has spent a decade on exactly that object under the name downbeat tracking. Study 36 is the proposal's first step, registered before any run: the published trackers against the production detector, graded at the bar, on every recording the program already holds truth for.

Systems and Truth

Production is the v0.200 detector port followed by the shipped grid, vote and placement, automatic mode exactly as the product would run it. Against it: Beat This! (ISMIR 2024, MIT weights) in its 78 MB and 8 MB checkpoints with the authors' minimal peak-picking, and the 78 MB model with madmom's dynamic Bayesian postprocessor; madmom's own RNN and DBN downbeat tracker with bars of three or four beats, and an exploratory variant that also allows six; librosa's unseeded tracker as the classical tempo baseline. Truth is the 23-recording union of Studies 34 and 35: five click-exact Badverbs mixes with uniform downbeat grids, three JAAH jazz 78s with beat-aligned downbeats and bar counts, the hymn and the waltz with tempo from the self-similarity instrument, ten Josh Woodward songs at the artist's declared tempo and meter (a 6/8 bar being two dotted quarters), and three meter-only slots. Graders are mechanical and unit-tested: bar period strict within 4 percent or folded at a metrical ratio, meter class, downbeat F-measure within 70 ms where time truth exists, bar counts on the 78s.

The Modern Ten

Bar period strict, folded, and meter right, of ten: production 3, 7, 8; Beat This final 6, 9, 8; Beat This small 7, 8, 8; madmom 7, 10, 8. The registered decision rule (strict on seven, folded on all ten, meter on eight) is met by madmom and by no Beat This variant, so the automatic-mode re-draft registered as Study 38 will run on madmom's bar. The misses are instructive. Every system reads goodbye-to-spring, declared 240, at 120, the registered counter-case that is a convention rather than an error. The small Beat This model's one clean failure is a three-beat bar on that same straight 4/4 song. And im-letting-go, declared 76, comes back at double tempo from every tracker, the other direction of the same convention.

The 78s and the Phase

On livery-stable-blues every neural tracker reads the 1.46 second bar that production reads as 2.71; on dippermouth 1.28 against 2.56; on st-louis 3.53 to 3.58 against 6.12. madmom places the downbeat within 70 ms on 96 percent of livery's bars and 66 percent of dippermouth's, with bar counts of 126 against 125 truth, 114 of 114 and 52 of 53, which is Study 34's exploratory tempo-following result delivered by a published system with no declared tempo given to it. Beat This gets the same periods but puts every downbeat on beat three of dippermouth's bars, a half-bar phase error that scores zero on downbeat F while the bars themselves are right; its DBN variant lands st-louis at 53 of 53. The hymn defeats every system, 2.30, 3.01 and 3.33 seconds against 2.47 declared, the recording the field's own literature names as the hard class. Beat This alone names the waltz's three; madmom reads four at half tempo.

Compound Meter Cuts the Other Way

Production's automatic mode gets gravity's 6/8 bar exactly: an eighth-note pulse of 145.9 with triplet subdivision becomes a 2.466 second bar against the true 2.5. Study 35's confidently wrong draft on that song was therefore the declared-tempo units bug alone, and the app's input adapter remains the whole fix there. The trackers are worse than the detector on this class. Beat This has no compound meter in its vocabulary and reads both 6/8 songs as three-beat bars, gravity as three eighths (half the bar) and my-favorite-regret as three beats of two per dotted quarter, a reading that folds to nothing. madmom with a six-beat class names both, then hears sixes in 4/4 jazz. If a tracker replaces the detector for the bar, the detector's subdivision read should keep the meter label.

Disclosures and Cost

On eat-all-the-clocks all five trackers place every downbeat 85 ms after the registered grid, uniformly, which is the registered bar-one offset itself; the 70 ms window makes that a zero for every tracker while production scores 0.99 from the same onset grid that produced the truth. Reported as registered; it is a fixture question and it treats every tracker alike. Prediction tally: P2, P3 and P5 right; P1 wrong on all three clauses (Beat This strict on six, folded on nine, compound on neither); P4 wrong as a conjunction through the dippermouth phase and the halved st-louis bar. Runtime on this VM was 27 to 47 seconds a song for the large Beat This model with the small one not reliably faster on CPU, because the spectrogram front end dominates; browser timing on WebGPU is a Study 38 matter. No model spend.